Retrieval Robustness

Retrieval robustness asks whether systems can find and use relevant evidence when wording, formats, modalities, collections, or evidence quality change, because retrieval failures can mislead downstream decisions and generated answers. Researchers test dual-encoder and late-interaction retrievers, multimodal search systems, and retrieval-augmented generation (RAG) pipelines, using methods such as adaptive reranking, evidence verification and abstention, representational stabilization, and targeted document augmentation. Evaluations span query rewrites and typos, distractors and collection growth, noisy image-text pairs, and poisoned or stale evidence, using ranking measures alongside answer accuracy, factuality, and attack success on tasks such as HotpotQA and Natural Questions. A central tension is that better clean-set rankings or poison detection do not guarantee reliable downstream answers: defenses may suppress legitimate evidence or add latency, so robustness needs to be measured across pipeline stages and conditions.